面向医学图像分割的超像素U-Net网络设计  

Design of Superpiexl U-Net Network for Medical Image Segmentation

作　　者：王海鸥 刘慧[1,2] 郭强 邓凯[3] 张彩明 Wang Haiou;Liu Hui;Guo Qiang;Deng Kai;Zhang Caiming(School of Computer Science and Technology,Shandong University of Finance and Economics,Ji’nan 250014;Digital Media Technology Key Laboratory of Shandong Province,Ji’nan 250014;Department of Image,Shandong Provincial Qianfoshan Hospital,Ji’nan 250014;School of Software,Shandong University,Ji’nan 250101;Shandong Co-Innovation Center of Future Intelligent Computing,Yantai 264025)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东省数字媒体技术重点实验室,济南250014 [3]山东省千佛山医院影像科,济南250014 [4]山东大学软件学院,济南250101 [5]山东高校未来智能计算协同创新中心,烟台264025

出　　处：《计算机辅助设计与图形学学报》2019年第6期1007-1017,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金重点项目浙江联合基金(U1609218);国家自然科学基金(61572286,61472220);山东省重点研发计划(2017CXGC1504);山东省省属高校优秀青年人才联合基金项目(ZR2017JL029);山东省高等学校优势学科人才团队培育计划

摘　　要：近年来,超像素在医学图像处理领域的应用愈加广泛,现有的方法取得了较好的效果,如LAW,SLIC等.然而,这些方法在处理医学图像得到超像素时,位于组织边缘像素点的划分仍存在类别模糊问题.为此,提出一种基于U-Net网络的超像素分割方法.首先,通过双边滤波模型过滤外部噪声,增强超像素信息;然后,结合U-Net卷积网络学习图像特征.该方法为U-Net网络中每个特征尺度的卷积层后嵌入一个规范层,用于增强网络对参数的敏感性.实验结果表明,该方法有效提高了医学图像超像素的分割精度,与groundtruth相比,其改善了超像素边缘分类的准确性,优化了超像素分割结果,在精确度、召回率、F-measure和分割速度等性能指标上均取得了更好的效果.In recent years, the superpixel methods have been widely used in the field of medical image processing and achieved good results, such as LAW, SLIC, etc. However, there were still some problems of fuzzy classification at the edge of the tissues when these methods were used to obtain superpixels. A superpixel optimization approach based on U-Net architecture was proposed in this paper. Firstly, a bilateral filtering (BF) operation was adopted to eliminate external noisy effects at the beginning of the network, and enhance the grayscale information. Then, via combining with U-Net networks, the whole model can learn the image features and output the optimized results for the superpixel map. In terms of network design, a normalization layer was embedded behind the convolution layer at each feature-scales, in order to strengthen the sensitivity of the parameters. Experimental results show that the classification accuracy in superpixel edge is significantly improved compared with the ground truth. Moreover, this method has achieved better results in precision, recall, F-measure and computational efficiency than other classic methods.

关 键 词：超像素 双边滤波 卷积网络 U-Net 医学图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]